D. Katabi, M. Handley, and C. Rohrs, “Congestion control for high bandwidth-delay product networks,” presented at the ACM SIGCOMM 2002, Pittsburgh, PA.
eXplicit Control Protocol (XCP) tries to address low performance of TCP over networks with high bandwidth-delay products. XCP builds on top of several earlier proposals such as Explicit Congestion Notification (ECN) and CSFQ to improve the link utilization and fairness in such networks. One idea used in this scheme is that implicit signaling is a poor signaling mechanism and with adding more bits to the feedback signal, XCP should be able to outperform TCP. Explicit signaling also has other advantages such as potentially easier policing mechanisms and therefore better security.
XCP architecture is constructed from XCP sender, receiver and router. In order to keep the per flow processing minimal at the routers, XCP sender attaches the window size and estimated rtt per flow to the packet which amazingly simplifies the per packet calculations needed at the router in order to achieve fairness per flow and high utilization.
One of the most interesting aspects of this architecture, is the separation of Efficiency Controller (EC) and Fairness Controller (FC). Linear fair queuing algorithms need to do AIMD in order to converge to both efficiency and fairness. The increase need to be linear or otherwise the algorithm wouldn't converge to fairness. By separating the EC and FC, EC can do MIMD and therefore converge much faster to efficient link usage while FC is still doing AIMD separately. This also enables the FC to implement other notions of fairness such as shadow pricing independent of EC.
The paper is fun to read and interesting. There were several places where I wasn't convinced enough with the arguments made for choosing parameters. For example the control window was chosen to be equal to average RTT time by following argument:
Even though simulations showed performance improvement over TCP/AQM schemes with certain simplistic topologies, without extensive implementation it is not easy to truely compare the schemes and understand the drawbacks of XCP approach. It is interesting to see if XCP has been further employed or tested over the last 7 years and if it has been widely deployed!
eXplicit Control Protocol (XCP) tries to address low performance of TCP over networks with high bandwidth-delay products. XCP builds on top of several earlier proposals such as Explicit Congestion Notification (ECN) and CSFQ to improve the link utilization and fairness in such networks. One idea used in this scheme is that implicit signaling is a poor signaling mechanism and with adding more bits to the feedback signal, XCP should be able to outperform TCP. Explicit signaling also has other advantages such as potentially easier policing mechanisms and therefore better security.
XCP architecture is constructed from XCP sender, receiver and router. In order to keep the per flow processing minimal at the routers, XCP sender attaches the window size and estimated rtt per flow to the packet which amazingly simplifies the per packet calculations needed at the router in order to achieve fairness per flow and high utilization.
One of the most interesting aspects of this architecture, is the separation of Efficiency Controller (EC) and Fairness Controller (FC). Linear fair queuing algorithms need to do AIMD in order to converge to both efficiency and fairness. The increase need to be linear or otherwise the algorithm wouldn't converge to fairness. By separating the EC and FC, EC can do MIMD and therefore converge much faster to efficient link usage while FC is still doing AIMD separately. This also enables the FC to implement other notions of fairness such as shadow pricing independent of EC.
The paper is fun to read and interesting. There were several places where I wasn't convinced enough with the arguments made for choosing parameters. For example the control window was chosen to be equal to average RTT time by following argument:
Estimating parameters over intervals longer than the average RTT leads to sluggish response, while estimating parameters over shorter intervals leads to erroneous estimates.It is not obvious why 1 average RTT is a better choice than 0.9 average RTT or 2 average RTT or if this makes any difference. As a matter of fact around 50% of effects of one control decision hasn't been observed yet after one average RTT.
Even though simulations showed performance improvement over TCP/AQM schemes with certain simplistic topologies, without extensive implementation it is not easy to truely compare the schemes and understand the drawbacks of XCP approach. It is interesting to see if XCP has been further employed or tested over the last 7 years and if it has been widely deployed!
No comments:
Post a Comment